Token Merging for Multilingual Speech Recognition: A Systematic Study Across Model Scale and Fine-Tuning

The work systematically evaluates inference-time token merging across a multilingual speech-recognition model family and examines its interaction with parameter-efficient adaptation in lower-resource settings.

Source: [14]

For the evaluated languages and checkpoints, the authors recommend a common high reduction setting without language-specific recalibration; lower settings offer additional accuracy margin when constraints are tighter.

Source: [18]

The study examines whether adjacent token merging transfers across model scales, linguistically diverse settings, and decoder-side adaptation.

Source: [9]

The evaluation draws from a multilingual speech benchmark with a language set spanning phonological variation, linguistic families, and resource strata.

Source: [7]

The study evaluates model variants spanning scale and applies decoder-only low-rank adaptation to the middle-scale variant while retaining its encoder.

Source: [8], [17]

The merging procedure scores adjacent representations by key-vector similarity, greedily forms disjoint pairs, and replaces selected pairs with averaged hidden representations.

Source: [6], [20]

Merging follows a fixed encoder-layer schedule that begins near the input and omits the final encoder layer to retain a buffer before decoder access.

Source: [16]

The evaluation compares an unmerged baseline with a range of token-reduction settings under a fixed merge-layer schedule.

Source: [13]

Recognition quality is assessed using normalized-transcript word error relative to the corresponding unmerged model, while efficiency is assessed through observed computation time.

Source: [3]

For the evaluated lower-resource cohort, the middle-scale model showed broadly stable error under merging, with the strongest language-level improvement occurring for a named member of that cohort.

Source: [4]

For the high-resource anchor cohort, merging produced only a small average error change under the stated evaluation setting.

Source: [10]

After decoder-side adaptation, merging retained a small average error effect on trained languages and a bounded largest degradation on held-out languages.

Source: [11]

Across the tested model scales, average error changes remained small, although the direction of the mean change differed by scale.

Source: [5], [12]

Latency measurements use device-timed paired baseline and merged runs, warm-up handling, synchronized execution, and separate encoder and end-to-end reporting in a fixed hardware and software environment.

Source: [15]

At the highest tested reduction setting, encoder latency declines relative to the unmerged baseline, yielding an encoder-local speedup whose deployment value is expected to depend on encoder cost.

Source: [2], [19]

The study uses a static merge schedule across languages and scales, and its adaptation-composition evidence is confined to the selected low-rank adaptation approach.

Source: [1]

The held-out evaluation covers only part of the pretraining distribution, so the findings support broad generalization without establishing exhaustive coverage.

Source: [1]